[llvm] [AArch64] Improve lowering of truncating uzp1 (PR #82457)

[David Sherwood via llvm-commits]

================
@@ -6105,26 +6105,39 @@ defm UZP2 : SIMDZipVector<0b101, "uzp2", AArch64uzp2>;
 defm ZIP1 : SIMDZipVector<0b011, "zip1", AArch64zip1>;
 defm ZIP2 : SIMDZipVector<0b111, "zip2", AArch64zip2>;
 
-def : Pat<(v16i8 (concat_vectors (v8i8 (trunc (v8i16 V128:$Vn))),
-                                 (v8i8 (trunc (v8i16 V128:$Vm))))),
-          (UZP1v16i8 V128:$Vn, V128:$Vm)>;
-def : Pat<(v8i16 (concat_vectors (v4i16 (trunc (v4i32 V128:$Vn))),
-                                 (v4i16 (trunc (v4i32 V128:$Vm))))),
-          (UZP1v8i16 V128:$Vn, V128:$Vm)>;
-def : Pat<(v4i32 (concat_vectors (v2i32 (trunc (v2i64 V128:$Vn))),
-                                 (v2i32 (trunc (v2i64 V128:$Vm))))),
-          (UZP1v4i32 V128:$Vn, V128:$Vm)>;
-// These are the same as above, with an optional assertzext node that can be
-// generated from fptoi lowering.
-def : Pat<(v16i8 (concat_vectors (v8i8 (assertzext (trunc (v8i16 V128:$Vn)))),
-                                 (v8i8 (assertzext (trunc (v8i16 V128:$Vm)))))),
-          (UZP1v16i8 V128:$Vn, V128:$Vm)>;
-def : Pat<(v8i16 (concat_vectors (v4i16 (assertzext (trunc (v4i32 V128:$Vn)))),
-                                 (v4i16 (assertzext (trunc (v4i32 V128:$Vm)))))),
-          (UZP1v8i16 V128:$Vn, V128:$Vm)>;
-def : Pat<(v4i32 (concat_vectors (v2i32 (assertzext (trunc (v2i64 V128:$Vn)))),
-                                 (v2i32 (assertzext (trunc (v2i64 V128:$Vm)))))),
-          (UZP1v4i32 V128:$Vn, V128:$Vm)>;
+def trunc_optional_assert_ext : PatFrags<(ops node:$op0),
+                                         [(trunc node:$op0),
+                                          (assertzext (trunc node:$op0)),
+                                          (assertsext (trunc node:$op0))]>;
+
+// concat_vectors(trunc(x), trunc(y)) -> uzp1(x, y)
+// concat_vectors(assertzext(trunc(x)), assertzext(trunc(y))) -> uzp1(x, y)
+// concat_vectors(assertsext(trunc(x)), assertsext(trunc(y))) -> uzp1(x, y)
+class concat_trunc_to_uzp1_pat<ValueType SrcTy, ValueType TruncTy, ValueType ConcatTy>
+  : Pat<(ConcatTy (concat_vectors (TruncTy (trunc_optional_assert_ext (SrcTy V128:$Vn))),
----------------
david-arm wrote:

This pattern is a subset of trunc_concat_trunc_to_xtn_uzp1_pat so it's not obvious to me which one would get picked. I think we normally use the AddedComplexity tablegen property to differentiate between cases like this - perhaps worth seeing if we need to use it here?

https://github.com/llvm/llvm-project/pull/82457